Visual Tracking Via Local Discriminative Reverse Sparserep Resentation

Visual target tracking is a fundamental topic in the field of computer vision, which plays a important role in video surveillance, motion recognition, traffic monitoring, military navigation and other applications. Over the past decade, researchers from all over the world have achieved breakthrough by exploiting the Sparse Representation theory, which has been successfully applied in the field of target tracking. Moreover, the Sparse Representation theory demonstrated to be feasible and effective.Target tracking methods based on sparse representation theory have achieved a major breakthrough, which has been applied to some challenging scenarios, such as pedestrian detection, vehicles tracking and other applications. However, the available algorithms still cannot perfectly solve the real-time tracking problems, such as the occlusion problem, the objective morphological change problem and complicated background problem. This thesis proposes two methods mainly to solve the occlusion problem and objective morphological change problem, and also verifies their effectiveness and robustness from theoretical and practical aspects.Firstly, an occlusion detection-based template updating method is proposed to solve the occlusion problem. The occlusion problem is intractable problems in target tracking field. The template set will be updated according to occlusion information of current frame. The template update method is in full use occlusion information by splitting the template into two subsets and picking one template of these two sets. Therefore, the discriminative information and accuracy of templates are improved.Secondly, the local discriminative reverse sparse representation model is proposed. The main focuses of the model are the occlusion problem and pose and scale change problem. Locality taking into accounts the major positions and minor positions. Discrimination considers the differences of foreground and background to identify the best candidate target.The experimental results have demonstrated that, the proposed model can effectively solve the occlusion problem and the objective morphological change problem. Furthermore, the proposed model possesses the capabilities of higher tracking accuracy and efficiency in comparison with other tracking algorithms.